// the find
kaixxx/noScribe
Cutting edge AI technology for automated audio transcription. A nice GUI for OpenAIs Whisper and pyannote (speaker identification)
noScribe is a desktop app that wraps faster-whisper and pyannote into a GUI aimed at qualitative researchers and journalists who need to transcribe interviews. It runs fully offline, handles speaker diarization, and ships a paired editor for post-transcription correction. This is squarely for academics and journalists, not developers building transcription pipelines.
The offline-first design is the real differentiator — no audio ever leaves your machine, which matters when the content is a sensitive interview. Pyannote integration for speaker diarization is bundled and working, not bolted on as an afterthought. The paired noScribeEdit app with audio-synced text playback is genuinely useful for review; most transcription tools dump a text file and leave you to figure out corrections. Batch queue support landed in 0.7, so multi-hour interview workflows are now first-class.
Intel Mac support is dropped in 0.7 with no clear timeline for a fix — half the MacOS install section is now pointing users to an older version. The download is several gigabytes because the AI models are bundled rather than fetched on first run, which makes distribution and updates clunky. Whisper hallucination on quiet passages is acknowledged but the VAD mitigation is imperfect, and hallucinated words that fit syntactically are specifically called out as hard to catch — that's a real problem for accuracy-critical research. The project is a solo maintainer with no governance structure, so the bus factor is exactly one.